Abstract: In allusion to recognition method for the red and white cells with vague boundary in the stool microscopy images, this paper proposes a new method based on Chan-Vese model, which considers both the edge and texture information of neighboring regions. It proposes a synergetic combination 8 direction Sobel edge enhancement with the tensor field as the region texture attribute in order to remedy the clear cell’s fuzzy boundary and keep the cell’s inherent texture. The complementary edge and texture is more suitable for the Chan-Vese segmentation model considering global and local energy distribution. It also takes the random decision forest with better data generalization effect as a classification tool. The experimental results show that the improved Chan-Vese segmentation method considering both boundary and domain texture makes the segmentation accuracy of red and white cells in the stool microscopy  images up to 95.3%. This method has higher recognition precision and strong universality for tangible object in stool microscopy images.

Key words: stool microscopy images, image segmentation, structure tensor, Chan-Vese model, random decision forest

摘要： 针对粪便镜检图像中具有弱边界的红、白细胞的识别问题，研究了基于Chan-Vese模型的兼顾邻域区域边缘和纹理综合信息的分割方法。用八向Sobel弥补透明细胞的模糊边缘，通过细胞域内纹理和边缘信息互补而采用兼顾全局和局部能量分布的Chan-Vese模型的分割方法，并采用具备更好的数据泛化作用的随机决策森林进行分类。实验证明，提出的兼顾边界与域内纹理的改进型Chan-Vese分割方法使粪便镜检图像中红、白细胞的分割精度达到了95.3%。该方法对粪便镜检图像中的有形物体具备更高的分辨能力和光学环境适应性。

关键词: 粪便镜检图像, 图像分割, 结构张量, Chan-Vese模型, 随机决策森林